Overview
- Anthropic acknowledged late Thursday that it ran an experiment beginning in March that inserted a hidden session marker into Claude Code to deter account resale and model ‘‘distillation’’.
- Researchers and developers reported the marker worked by checking users' time zones and proxy settings and then making tiny, hard-to-see text changes that could identify flagged sessions.
- Anthropic says it planned the marker's rollback, has already built stronger protections, is tightening identity verification and anti-proxy signals, and is cooperating with U.S. officials on model safety.
- Several major organisations have restricted Claude access, with Alibaba ordering an internal ban on Claude Code for workplace use starting July 10 and other banks and tech firms imposing limits.
- The episode has deepened tensions over cloud-hosted model control, pitting model-protection tools against demands for transparency and likely accelerating enterprise moves to domestic or better-audited AI stacks.